IACBAM-YOLO: an enhanced lightweight YOLO for kiwifruit detection under complex lighting and occlusion conditions

Accurate identification of kiwifruit in unstructured orchard is a prerequisite for achieving automated robotic harvesting. Variable lighting conditions, as well as overlapping foliage, branches, and other fruits, reduce the distinguishable features of kiwifruit in images, thereby increasing the difficulty of accurate detection. The You Only Look Once (YOLO) series of models has seen widely adopted for object detection due to its strong balance between speed and precision. However, existing YOLO-based algorithms still face challenges in maintaining detecting performance under complex lighting conditions and target occlusion. To address these issues, this study introduces an improved YOLOv8n algorithm based on a light-sensing adaptive Illumination-aware Adaptive Convolutional Block Attention Module (IACBAM). The environmental-awareness branch dynamically adjusts the channel and spatial attention weights on the IACBAM by generating attention-scaling coefficients based on current lighting conditions. Integrated into key layers of the YOLO network, the IACBAM enhances fruit-related features while suppressing background interference. An end-to-end joint optimization training strategy is employed to enable collaborative learning and unified training of the YOLO backbone, the CBAM, and the illumination-aware branch. Experimental results using a self-constructed dataset demonstrate that the proposed algorithm achieves precision of 90.2%, recall of 84.7%, and mAP50 of 93.7%. Compared to Faster R-CNN, YOLOv3‑tiny, and the original YOLOv8n, this method significantly enhances the robustness of kiwifruit detection under challenging lighting and occlusion conditions while maintaining the lightweight characteristics of the network. Finally, the proposed approach provides a technical reference for fruit recognition in robotic by harvesting systems operating in unstructured orchards.